Explainable Machine Learning for Characterizing Unknown Molecular Structures in Infrared Spectra.

Area of Science:

• Chemistry
• Spectroscopy
• Machine Learning

Background:

• Infrared (IR) spectrum analysis is crucial for identifying molecular functional groups but is often time-consuming and labor-intensive.
• Existing machine learning methods for IR spectral analysis lack the incorporation of prior chemical knowledge and suffer from a lack of explainability due to their black-box nature.

Purpose of the Study:

• To develop an efficient and explainable deep learning method for functional group detection in IR spectra of unknown molecules.
• To address the limitations of existing machine learning approaches by integrating prior knowledge and enhancing model interpretability.

Main Methods:

• Proposing the Substructure-Directed Spectrum Interpreter Network (SSIN), a novel deep learning architecture.
• Training and evaluating SSIN on a large dataset of experimentally measured gas-phase IR spectra from the NIST database.
• Utilizing large language models to generate human-readable IR spectrum analysis reports.

Main Results:

• SSIN achieved a detection accuracy greater than 0.920 on the NIST database (8845 spectra).
• The model accurately identified specific absorption peaks corresponding to target functional groups.
• Generated IR spectrum analysis reports showed 81-99% accuracy compared to ground-truth data.

Conclusions:

• SSIN offers an efficient and explainable solution for functional group detection in IR spectroscopy.
• The integration of prior knowledge and explainability in SSIN overcomes key limitations of previous machine learning methods.
• Publicly available source codes and models facilitate further research and application.